2.1 Corruption in Latin America

Our article focuses on grand corruption in the public sector, a form of corrupt behavior that typically involves fewer people but larger exchanges of money (Prasad et al. Reference Prasad, Martins da Silva and Nickow2019). Focusing on grand corruption allows us to assess the determinants of electoral penalization, given that it is more pertinent to high-level politics, as opposed to bureaucratic or petty corruption. We narrow the conceptualization of corruption by focusing on the misappropriation of funds, using positions of power in political spheres for private gain. It is important to make this distinction given that petty corruption does not tend to be a major focal point for media reports (given its broader systematic and cultural nature). On the other hand, media-exposed cases of public sector corruption in the form of funds misappropriation typically relate to the behavior of specific culprits (politicians in our case).

Public sector corruption in Latin America consistently ranks as one of the highest in the world. Except for Chile and Uruguay, all Latin American countries rate poorly in corruption indices. Transparency International rates the Latin American region with a Corruption Perception Index (CPI) of 35 on average compared to an average global CPI of 43 points, meaning that Latin America has a higher corruption level than the world average (Transparency International 2021). This is corroborated by the V-Dem’s index on corruption in the public sector (V-Dem 2021). Corruption at such high levels entails significant developmental costs. It negatively affects economic growth, investment, and government expenditure (Mauro Reference Mauro1995). This has been corroborated by a large number of studies (Shaw et al. Reference Shaw, Katsaiti and Jurgilas2011; d’Agostino et al. Reference d’Agostino, Paul Dunne and Pieroni2016; de Paulo et al. Reference de Paulo, de Andrade Lima and Tigre2022; Everhart et al. Reference Everhart, Martínez-Vázquez and McNab2009; Sharma and Mitra Reference Sharma and Mitra2019). Additionally, corruption affects many other developmental factors, as in the case of mortality rates (Escaleras and Register Reference Escaleras and Register2016), environmental conservation (Pellegrini Reference Pellegrini2011), health (Anbarci et al. Reference Anbarci, Escaleras and Register2009) and education (Heyneman et al. Reference Heyneman, Anderson and Nuraliyeva2008; Duerrenberger and Warning Reference Duerrenberger and Warning2018).

Data from a 2018 survey by CAF shows that Latin Americans are aware of the harm corruption entails for economic progress, social equality, and the quality of public services. They even single out corruption as the main issue of concern their countries face. In light of this, some turn to electoral penalization as a viable anti-corruption solution. However, Latin Americans have in large part demonstrated a striking lack of penalization towards corrupt politicians in elections (Berniell et al. Reference Lucila, de la Mata and Italia2019). Scholars have attempted to understand this phenomenon. In some cases, voters expect the benefits of certain politicians being in office to exceed the costs of their corrupt acts (Winters and Weitz-Shapiro Reference Winters Matthew and Weitz-Shapiro2013). In other instances, voters living in countries with weak and corrupt institutions feel resignation towards political institutions and do not think their vote will change the state of corruption in the public sector (Agerberg Reference Agerberg2019). Most scholars, however, seem to argue that Latin Americans fail to penalize corrupt politicians due to lack of information in the media. The effect of the media (or information availability) on voter penalization has been confirmed by several seminal studies. Ferraz and Finan (Reference Ferraz and Finan2008) demonstrate that the availability of information from state audits increases electoral penalization of corrupt politicians in Brazil, while the propagation of these audits in the media increases this effect. Bobonis et al. (Reference Bobonis, Cámara Fuertes and Schwabe2016) follow this line by showing that Puerto Rican municipalities with timely audits before elections have considerably lower levels of corruption. The dissemination of audit results is paramount for these effects to materialize. Additionally, a study on Mexico specifies that the availability of information about corrupt incumbents does increase electoral accountability, but it decreases voter turnout and does not increase the support for challengers’ parties (Chong et al. Reference Chong, Ana, Karlan and Wantchekon2012). In addition, recent studies find a negative relationship between corruption and social media penetration, even in contexts of restricted press freedom (Enikolopov et al. Reference Enikolopov, Petrova and Sonin2018; Jha and Sarangi Reference Jha and Sarangi2017). The higher the social media use is, the lower the metrics of corruption. Furthermore, highly connected networks enable voters to coordinate among themselves and share information to penalize corrupt politicians (Arias et al. Reference Arias, Balán, Larreguy, Marshall and Querubín2019). Hence, the connectivity and openness that social media offers may indicate a newfound opportunity to curb corruption.

2.2 Media Credibility

Credibility refers to the expectation that someone’s word can be relied on. This expectation is built on that person’s reputation, which depends on consistently providing truthful information (Sobel Reference Sobel1985, 557). The study of credibility in media can be categorized into two areas (see Golan Reference Golan2010, 10): studies on medium credibility (i.e., media channels such as television, newspapers, etc.) and studies on source credibility (i.e., the messenger such as a speaker, organization, author, etc.). In the case of Latin America, social media is the second most popular channel of information about politics (CAF 2018). Tsfati (Reference Tsfati2010, 38) claims that the rising tendency of alternative media outlets is driven by heightened skepticism towards traditional media. However, non-mainstream media often have defining characteristics that tend to clash with journalistic professionalism or traditional news values (Tsfati Reference Tsfati2010, 26). It is no secret that information posted online often does not entail an exhaustive factual verification (Flanagin and Metzger Reference Flanagin and Metzger2000, 516). Given this, several studies focus on social media credibility topics such as fact checking, evaluating indicators of credibility, curbing the spread of misinformation, and the recent phenomenon of fake news (Viviani and Pasi Reference Viviani and Pasi2017; Yaqub et al. Reference Yaqub, Kakhidze, Brockman, Memon and Patil2020). On the contrary, in the case of traditional media, Asak and Molale (Reference Asak and Molale2020) demonstrate that newspapers have low instances of fake news and generally abide by professional journalism principles. The literature presents a scholarly debate on whether social media is deemed more (or less) credible than traditional sources such as newspapers or television broadcasts. On the one hand, Karlsen and Aalberg (Reference Karlsen and Aalberg2021) propose that news published on social media are deemed less credible by audiences, particularly when politicians are intermediary messengers. On the other hand, Johnson and Kaye (Reference Johnson and Kaye2010) claim that audiences consider the internet more trustworthy than traditional media, particularly during electoral and campaigning seasons. In addition to the study of medium credibility, scholars argue that the source of information also plays a key role in credibility. Miller and Kurpius (Reference Miller and Kurpius2010) assessed the determinants of credibility based on the characteristics of messengers in television broadcasts; Meyer et al. (Reference Meyer, Marchionni and Thorson2010) suggest that it is the author’s perceived expertise that influences the perceived credibility by news audiences. In the case of social media, Tandoc (Reference Tandoc2019) claims that the credibility attributed to news depends on the messenger (an organization vs. friends), particularly when the motivation to engage with the news is high. If motivation is low, audiences tend to engage in more heuristic cues that affect their ability to discern credibility in sources. Similarly, other studies assessed how heuristics related to the source and the recency of the post affect the perception of credibility on social media (Lin et al. Reference Lin, Spence and Lachlan2016; Westerman et al. Reference Westerman, Spence and van der Heide2014).

3.1 The Experiment

One particular experiment from ECAF 2018 motivates this study. This experiment can be found in question 43 in the ECAF 2018 questionnaire. ^{Footnote 4} In this experiment the respondent is told about a report exposing irregularities in the purchase of materials authorized by a government minister. This report is published in different media channels/sources, which are considered treatments in the experiment. The respondents are thus randomized across three treatment groups and one control group. The control group is told that a state auditor releases the report. The first treatment group is told that the nation’s most popular newspaper releases the report. The second treatment group is told that a named journalist publishes the report on social media (Facebook, Twitter, Instagram, etc.). The third treatment group is told that a group of anonymous journalists publish the report on social media. The researcher then asks two questions:

The experimental data comprise 7,558 observations for the credibility question and 7,561 observations for the penalization question. ^{Footnote 5} It should be noted that the experiment is subject to some limitations. One of the limitations arises from the fact that this is an experiment embedded in a survey. This may raise concerns about the survey questions prior to the experiment, given that they may condition the responses to the treatments. As the survey proceeds, the respondent’s awareness of corruption may become more salient and thus increase one’s inclination to believe a report about a corrupt politician regardless of the media source. In other words, one’s perception of credibility may be overestimated. Similarly, one’s willingness to penalize a corrupt politician might increase throughout the survey. However, while this may affect the overall mean of the outcome variables, it should not affect the differences between treatments. Another limitation is that the treatments in the experiment are not made salient in any way to the respondents. The report and government minister that are central to the questions in the experiment are both hypothetical and immaterial. While this ensures that no political or media figure is attached to the conceptualization of the experiment, this does leave the concept of the corruption report open to interpretation. The article hence makes the assumption that the corruption report is conceptualized similarly by respondents (in other words, that the understanding of the treatment is homogenous across all respondents). Additionally, the article assumes that respondents do not make associations between the hypothetical government minister and relevant political figures in their respective countries.

The article hence utilizes three treatment groups and one control group. The control group includes the state as a source of information to assess how much people trust their governments compared to other communication outlets (i.e., to contrast credibility in journalistic vs. non-journalistic government information). The first and second treatment groups include traditional and social media channels (in both cases with information provided by named journalists) to address the implications of the rise of social media, which is relevant to the first and second hypotheses of the article. Finally, the third treatment involves anonymous journalists on social media, which is crucial for two reasons. First, the inclusion of this treatment is relevant because it can indicate the level of attention audiences dedicate to credibility cues (see Section 2.2). There are two possible opposing forces here. On the one hand, anonymous social media postings may be associated with reduced credibility as this prevents identifying the identity of the sender and verifying the accuracy of information. On the other hand, anonymous postings may be considered more credible since this allows journalists to express their opinions more freely and without the risk of external interferences or retribution. Second, this third treatment is likely to be particularly relevant in environments (not atypical in many Latin American contexts) where freedom of expression is insufficiently protected and when journalists who investigate and report on corruption cases become harassed and threatened (Pennachio Reference Pennachio2021; Nalvarte Reference Nalvarte2020; Aruth Sturm Reference Aruth Sturm2016); this hence incentivizes some journalists to publish their investigations anonymously. As a matter of fact, several Latin American journalists make use of anonymous blogging platforms (such as telegra.ph or write.as) or post in more common social media platforms (e.g. Facebook) with the use of pseudonyms.

In order to facilitate the interpretation of results and the analysis in general, the variables for credibility and penalization were both redefined into binary variables. Instead of carrying out the study on a Likert scale, the two outcome variables of interest are stipulated in binary terms. The definition of these two variables can be seen in Table 1 above.

Table 1. Binary Variables for Credibility and Penalization

Source: Authors’ analysis using data from ECAF 2018 (CAF 2018).

3.2 ECAF 2018 Data

The analysis of the experiment is complemented with additional variables from ECAF 2018. The purpose of this data is to analyze certain individual characteristics that could influence the respondents’ perception of credibility and their willingness to penalize corruption within each treatment group. Here we focus on three key determinants (variables).

The first determinant is the maximum level of education attained by the respondent. The survey’s original variable includes 13 levels of education. For the purpose of our analysis, this is simplified into a binary variable distinguishing between those who did not complete secondary school versus those who attained an educational level of secondary school and above. This distinction does not only serve as a simplification for the analysis, but it is also pertinent to the context of Latin America. According to World Bank Data (2020), the average rate of people who completed secondary school in the countries relevant to this study is 49 percent. Hence, attaining a secondary school qualification marks a balanced distinction between those who are better educated from those who are not so well educated. As Weitz-Shapiro and Winters (Reference Weitz-Shapiro and Winters2017) demonstrated, education plays an important role in people’s ability to discern credible sources. It is also linked to their level of political sophistication.

The second determinant of interest relates to a minimal level of political sophistication. This is measured in the survey by asking if the respondent is familiar with concepts of right-wing and left-wing politics. The addition of this variable allows us to test Winters and Weitz-Shapiro’s theory of political sophistication and its effect on credibility. It should be noted that there are some missing values for the political sophistication variable (which, however, only amount to 2 percent of the total observations and hence pose no major concern).

The third determinant involves the respondent’s perception of his or her fellow citizens’ concern about corruption. In other words, the variable responds to the question, “Do you believe your fellow citizens consider corruption events when casting their vote?” (CAF 2019). This variable aims to indicate the perception respondents have about the public’s general attitude towards corruption in their country (and its possible penalization by the average voter at the ballot box).

Table 2 presents the descriptive statistics for all variables of our analysis (The Supplementary Material Section presents the corresponding subgroup city values).

Table 2. Descriptive Statistics

Source: Authors’ analysis using data from ECAF 2018 (CAF 2018).

4.1 Average Treatment Effects

We first obtain the Average Treatment Effect (ATE) per outcome to evaluate the differences in treatment groups and the control group (regarding perceived credibility and penalization intent). Since the sample and the treatment assignation are randomized, we assume that the Average Treatment Effect on the Treated (ATT) is equal to the ATE. In other words, the selection and heterogeneity effects in the model should not be a concern. Also, spillover effects from treatments are not a concern since the treatments are assigned instantaneously in the household survey. There is no time gap between the assignment and the response to the treatment for any spillover effects to materialize.

There are two models, one for each outcome variable:

$${\rm{Credibilit}}{{\rm{y}}_{{\rm{ij}}}} = {\rm \alpha} + {\rm \beta _1}{{\rm{T}}_1} + {\rm \beta _2}{{\rm{T}}_2} + {\rm \beta _3}{{\rm{T}}_3} + {\rm \delta _{ij}} + {\rm \varepsilon _{{\rm{ij}}}}$$

$${\rm{Penalization}_{{\rm{ij}}}} = {\rm \rho} + {\rm \gamma _1}{{\rm{T}}_1} + {\rm\gamma _2}{{\rm{T}}_2} + {{\rm \gamma} _3}{{\rm{T}}_3} + {\rm \eta _{{\rm{ij}}}} + {{\rm\xi} _{{\rm{ij}}}}$$

where i and j refer to individuals and countries respectively. The coefficients of interest in these models are those pertaining to the treatment dummy variables (T₁, T₂, T₃). Each of these are compared to the control group, which acts as a reference category. The coefficients thus provide the treatment effects for every treatment group. The models also include constants α/ρ, country fixed effects δ/η, and error terms ϵ/ξ.

4.2 Heterogeneous Treatment Effects

Our analysis defines specific heterogeneous groups to assess whether they incur an additional effect on credibility or penalization based on their treatment. In this case, the article assesses whether there is a heterogeneity effect in the ATE tested above. For this purpose, we extend the aforementioned specifications by including interaction terms between each treatment group and a vector of person-specific variables X_ij capturing the relevant characteristic per heterogeneous group (T_kX_ij), as well as the X_ij variables alone. The coefficient of interest in this case is the one pertaining to the interaction term, which provides the additional effect of the group variable on the treatment effect. This specification allows us to assess whether certain characteristics of respondents enhance or diminish treatment effects on credibility and/or penalization.